Ensembles of data-reduction-based classifiers for distributed learning from very large data sets

An ensemble of classifiers is a set of classification models and a method to combine their predictions into a joint decision. They were primarily devised to improve classification accuracies over individual classifiers. However, the growing need for learning from very large data sets has opened new application areas for this strategy. According to this approach, new ensembles of classifiers have been addressed to partition a large data set into possible disjoint moderate-sized (sub)sets, which are then distributed along with a number of individual classifiers across multiple processors for parallel operation. A second benefit of reducing sizes is to make feasible the use of well-known learning methods which are not appropriate for handling huge amount of data. This paper presents such a distributed model as a solution to the problem of learning on very large data sets. As individual classifiers, data-reduction-based techniques are proposed due to their abilities to reduce complexities and, in much cases, error rates.